
Negative control outcomes for
empirical calibration
• Observational data analyses may have residual bias, so it’s important to
perform diagnostics to quantify the extent of this potential issue
• Bias = expected value of the error distribution (random + systematic)
• Negative control outcomes can be used efficiently in cohort analyses
– Outcomes which have no evidence about association with either target cohort
or outcome cohort, therefore ‘true RR’ assumed to equal 1 and any difference
between effect estimate and ‘true RR’ can be classified as systematic error
– Convention: find outcomes there ‘absence of evidence’ can be inferred to be
‘evidence of absence’:
1. not listed on target/comparator product labels
2. not co-occurring with target/comparators in published literature (Medline)
3. don’t have increased signal score from spontaneous adverse event reporting (FAERS)
4. do appear with adequate prevalence in the observational database so that an effect could
have been previously observable had it existed
• Sample of negative control outcomes (n>20) can be used to estimate
‘empirical null’ distribution, which can then be used to empirically
calibrate p-value for unknown outcome of interest